Jailbreak Attack on Harmful Query Evaluation Set N=750 GPT-5.4-nano
1.05Toxicity ScoreHADES
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| HADESNumber of Trials (T)=12026.05 | 1.05 | 0.88 | 1.33 | 1.73 | |
| FigStepNumber of Trials (T)=12026.05 | 1.11 | 1.23 | 3.2 | 3.33 | |
| FlipAttackNumber of Trials (T)=12026.05 | 1.16 | 0.72 | 4.8 | 5.6 | |
| SINumber of Trials (T)=52026.05 | 1.17 | 1.67 | 4.27 | 5.07 | |
| CS-DJNumber of Trials (T)=12026.05 | 1.86 | 2.75 | 25.73 | 14.4 | |
| TxtVNumber of Trials (T)=52026.05 | 2.62 | 4.26 | 42.8 | 33.73 | |
| TypVNumber of Trials (T)=52026.05 | 3.38 | 5.64 | 61.07 | 54 | |
| TxtV-GDINumber of Trials (T)=52026.05 | 3.65 | 6.15 | 69.6 | 62.4 | |
| TTVNumber of Trials (T)=52026.05 | 3.75 | 6.42 | 71.47 | 68.93 | |
| TTV-GDINumber of Trials (T)=5, Kd=62026.05 | 4.03 | 6.94 | 78.67 | 74.67 |